multi-way pipelining for power-efficient ip lookup
DESCRIPTION
MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP. Author: Weirong Jiang and Viktor K. Prasanna Publisher/Conf .: IEEE GLOBECOM 2008 Speaker: Han-Jhen Guo Date: 2008.12.17. OUTLINE. Background and Related Work Proposed Scheme introduction of proposed scheme - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/1.jpg)
MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP
Author: Weirong Jiang and Viktor K. Prasanna
Publisher/Conf.: IEEE GLOBECOM 2008
Speaker: Han-Jhen Guo
Date: 2008.12.17
1
![Page 2: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/2.jpg)
OUTLINE
Background and Related Work Proposed Scheme
introduction of proposed scheme hybrid SRAM/TCAM-based partitioning fine-grained node-to-stage mapping
Performance
2
![Page 3: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/3.jpg)
BACKGROUND AND RELATED WORK
Uni-bit trie
3
Leaf pushing
![Page 4: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/4.jpg)
Proposed Scheme- Introduction Of Proposed Scheme
Multi-way linear pipeline architecture (P=4)
each pipeline stores a portion of the routing table hybrid SRAM/TCAM-based partitioning scheme
ensure that pipelines contain exactly equal numbers of trie nodes
fine-grained node-to-stage mapping scheme within each pipeline balance the memory distribution over pipeline stages 4
![Page 5: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/5.jpg)
Proposed Scheme- Hybrid SRAM/TCAM-based Partitioning
1. Prefix expansion
initial stride larger, smaller subtries → can help balance the memory
distribution across stages when mapping nodes to stages within a pipeline
more prefix duplication (a prefix may be copied to multiple subtries)
5
initial stride (I)
![Page 6: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/6.jpg)
Proposed Scheme- Hybrid SRAM/TCAM-based Partitioning
2. Post-order split ensures that each pipeline contains ceiling of
(total nodes of subtries / number of pipelines) nodes (except the last one pipeline) sort the subtries obtained after prefix expansion in
decreasing order of size
6
ceiling of (3+11+1+3) / 2
![Page 7: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/7.jpg)
Proposed Scheme- hybrid SRAM/TCAM-based partitioning
3. Index table implementation
7
based on I
remaining subtries
priority(SRAM_B) > priority(SRAM_A)
![Page 8: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/8.jpg)
Proposed Scheme- fine-grained node-to-stage mapping
1) Leaf reduction
8
subtries in pipeline 1
after leaf reduction
![Page 9: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/9.jpg)
Proposed Scheme- fine-grained node-to-stage mapping
Eg. of node-to-stage mapping with pipeline 1
9
![Page 10: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/10.jpg)
Proposed Scheme- hybrid SRAM/TCAM-based partitioning
Index table implementation
10
address of the subtrie’s root in the 1st stage of pipeline
0
1
2
…
…
…
…
0
1
2
![Page 11: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/11.jpg)
PERFORMANCE
Node distribution over 8 pipelines results for 16 routing tables
11
partitioning scheme achieved a balanced memory allocation among the eight pipelines
![Page 12: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/12.jpg)
PERFORMANCE
Node distribution over 25 stages in the 8-way pipeline architecture
12
there were few nodes at the top levels of any subtrie→ unbalanced at first several stages each pipeline
![Page 13: MULTI-WAY PIPELINING FOR POWER-EFfiCIENT IP LOOKUP](https://reader036.vdocuments.mx/reader036/viewer/2022062518/56814000550346895dab3563/html5/thumbnails/13.jpg)
13